Shut down cleanly

How to do it

Handle SIGTERM. Fail the readiness check first and wait about 5 seconds, so the load balancer stops sending traffic. Then close the server and wait for open requests. Keep the total below the platform's kill timeout, such as 25 of Kubernetes' default 30 seconds. Then close the database. Without this, every deploy restart drops requests in flight and visitors see errors such as 502.

Example

process.on('SIGTERM', () => {
  ready = false;
  setTimeout(() => server.close(async () => { await db.close(); process.exit(0); }), 5_000);
  setTimeout(() => process.exit(1), 25_000).unref();
});

// Go: stop on SIGTERM, give requests in flight 20 seconds, then exit
ctx, stop := signal.NotifyContext(context.Background(), syscall.SIGTERM)
defer stop()
go server.ListenAndServe()
<-ctx.Done()
ready.Store(false)
time.Sleep(5 * time.Second)
shutdown, cancel := context.WithTimeout(context.Background(), 20*time.Second)
defer cancel()
server.Shutdown(shutdown)

# Python with an ASGI server such as uvicorn: it finishes open requests on SIGTERM.
# Close your own resources in the shutdown hook.
from contextlib import asynccontextmanager

@asynccontextmanager
async def lifespan(app):
    yield
    await database.close()

How to check

Start a slow request, then stop the service. The request finishes, and the logs show a clean exit within the time limit.